AI has the potential to lift long‑run growth materially, yet bottlenecks in human judgment, coordination and institutions mean gains may be delayed and vulnerabilities concentrated, so safety and governance must be advanced alongside automation.

Main Finding

Automating intelligence (AI for cognitive tasks, and AI plus robots for physical tasks) could plausibly raise long‑run economic growth rates above the historical US norm (~2% per year). However, realization of those gains is likely to be delayed by “weak links” — important tasks or decisions humans still must perform. Paradoxically, while weak links slow aggregate gains, they can make the system more vulnerable to concentrated harms because breaking a single critical link can cause large damage (e.g., misuse of superhuman software engineering).

Key Points

• Historical baseline: US growth has averaged roughly 2%/year for 150 years, but that stable rate does not imply AI cannot accelerate growth.
• Three core claims of the essay:
• Full automation of human tasks could materially increase growth rates.
• Weak links (tasks humans still need to perform, coordination frictions, institutional and cognitive bottlenecks) will slow the pace at which automation raises output; having much faster components (e.g., computation) does not automatically translate to proportional productivity gains if key complementary tasks remain manual.
• Weak links increase systemic risk: when performance hinges on the weakest element, damaging or corrupting one element (e.g., a powerful AI used for malicious hacking or bioengineering) can have outsized harms.
• Example intuition: although transistor counts and raw compute have exploded, productivity did not increase by the same factor because people still choose how to apply computation — identifying the right problems, interpreting results, making strategic decisions.
• Timing matters: because many weak links must be automated before large macro gains appear, growth acceleration could be delayed; but malicious actors may exploit narrow superhuman capabilities earlier, speeding up certain risks relative to benefits.

Data & Methods

• Empirical anchor: observation of long-run US growth ≈ 2%/year over ~150 years.
• Methodological approach: conceptual/theoretical analysis using a “weak-link” framework (the productivity of a process is limited by its least-automated or weakest component).
• Uses thought experiments and illustrative examples (transistor progress vs human productivity, software engineering and bio-risk scenarios) rather than new econometric estimates or microdata.
• No detailed calibration or causal identification is provided in the essay; arguments are qualitative and model‑based, aimed at plausibility and mechanism identification rather than precise prediction.
• Implied modeling directions: formal growth models with complementary tasks and bottlenecks, and risk models where single-link failures have outsized consequences.

Implications for AI Economics

• Growth prospects:
  • AI could be a powerful engine of faster aggregate growth if it can automate the majority of weak links across production, decision-making, and coordination.
  • Uncertainty about timing: large gains may be lumpy and delayed until many bottlenecks are resolved.
• Policy and research priorities:
  • Focus on identifying and automating critical weak links (complementary R&D) to translate compute advances into aggregate productivity.
  • Simultaneously invest heavily in safety, governance, and resilience: weak links create concentrated attack surfaces that can be exploited before broad economic gains materialize.
  • Strengthen institutions, monitoring, and contingency planning for high‑impact domains (financial systems, biotech, critical infrastructure).
• Labor and distributional effects:
  • The path of automation (which weak links get automated first) will shape labor demand, sectoral transitions, and inequality — anticipating bottlenecks can guide retraining and social policies.
• Research agenda for AI economics:
  • Empirically map weak links across sectors and estimate how automating each affects productivity.
  • Build formal growth models incorporating complementarities, bottlenecks, and endogenous investment in automating weak links.
  • Quantify risk channels where narrow superhuman capabilities can be misused and evaluate mitigation strategies (regulation, secure design, access controls).
• International coordination:
  • Because misuse can be global and concentrated, cross‑border cooperation on safety standards, information sharing, and norms is important to manage downside risks while capturing upside growth.